Could AI Help Patients Get Off IV Antibiotics Sooner?

Switching patients from intravenous (IV) to oral antibiotics as soon as it's medically appropriate is one of those small-sounding decisions that actually has a big impact on hospital care. Staying on an IV line longer than necessary can raise the risk of catheter-related infections, keep patients in the hospital longer, and drive up healthcare costs. Yet according to a new study published in Nature Communications, roughly one in five patients in England remain on IV antibiotics even after they meet the clinical criteria to switch to oral treatment. Researchers from University College London and University College Hospital London set out to see whether a smarter kind of computer decision-support tool could help close that gap.

Why the Delay Happens

Deciding when to switch a patient from IV to oral antibiotics involves tracking multiple vital signs, like temperature, heart rate, and blood pressure, against established clinical guidelines. In the busyness of hospital wards, these opportunities can be easily missed or delayed. Some hospitals already use computerized decision-support systems to flag switch-ready patients, but the authors point out a subtle flaw in many of these tools: they're often trained on historical hospital data reflecting what clinicians actually did in the past, rather than what guidelines recommend. That means these systems can end up learning and reinforcing the very delays and inconsistencies they're supposed to fix.

Forecasting Instead of Mimicking Past Behavior

Instead of training a model to copy past clinical behavior, the research team designed a system built around forecasting. Using a machine learning technique called neural processes, the model predicts how a patient's vital signs are likely to trend over time, then compares these predicted trajectories directly against established clinical switching guidelines rather than against what doctors typically did in similar past cases. Anchoring predictions to guidelines rather than historical practice allows the system to flag patients who are genuinely ready for a switch, even in cases where routine past care might have overlooked them.

The tool doesn't make the final call itself; instead, it ranks patients to help clinical teams prioritize who to review first, keeping human judgment central to the decision-making process. The researchers also designed the system so it can be updated if clinical guidelines change, without needing to retrain the whole model from scratch, an appealing feature for real-world hospital adoption.

How the System Was Tested

The team evaluated their approach using two large, real-world datasets: over 6,300 patient encounters from US intensive care units and more than 10,500 encounters from a UK academic hospital group. Compared to randomly selecting patients for review, the forecasting-based system was 2.2 to 3.2 times more effective at identifying patients who were genuinely appropriate candidates for switching therapy. The authors also emphasize that the system's outputs are interpretable, meaning clinicians can potentially understand why a particular patient was flagged, rather than treating the tool as an opaque "black box."

A Proof of Concept, Not a Hospital Rollout

This is a computational and validation study, not a clinical trial. The researchers tested the tool against historical datasets to see how well it could identify switch-ready patients; they did not deploy it in real hospital wards to measure actual outcomes like length of stay or infection rates. The results show a strong association between the model's predictions and guideline-based readiness criteria, which is a meaningful proof of concept, though it stops short of demonstrating that using the system would directly improve patient outcomes in practice.

Even with that caveat, the work points to a useful shift in how decision-support tools get built: grounding predictions in clinical guidelines and physiological forecasting, rather than learning from imperfect past behavior. Should future prospective studies bear this out, tools like this could eventually support antibiotic stewardship efforts, helping hospital teams work more efficiently while keeping clinicians firmly in charge of final decisions. For now, questions about antibiotic treatment or hospital care decisions are best directed to the treating physician or care team involved.

Source: Ross, M., Swanepoel, N., Luintel, A. et al. "Optimising antibiotic switching via forecasting of patient physiology." Nature Communications (2026). DOI: 10.1038/s41467-026-76715-w. This post summarizes findings reported by the study's authors, based at UCL and University College Hospital London.